Monitoring and Fault Detection with Multivariate Statistical Process Control (MSPC) in Continuous and Batch Processes Barry M. Wise, Ph.D. Eigenvector Research, Inc. Manson, WA
1
Outline • Definition of Chemometrics • Favorite tools • Principal Components Analysis (PCA) • Partial Least Squares Regression (PLS) • Multi-way methods
• Opportunities in PAT • • • •
2
Multivariate Statistical Process Control (MSPC) Image analysis on tablets Predicting monitored or controlled variables Batch MSPC
Chemometrics Chemometrics is the chemical discipline that uses mathematical and statistical methods to 1) relate measurements made on a chemical system to the state of the system, and 2) design or select optimal measurement procedures and experiments. 3
Multivariate Analysis Multivariate Statistical Analysis is concerned with data that consists of multiple measurements on a number of individuals, objects, or data samples. The measurement and analysis of dependence between variables is fundamental to multivariate analysis. 4
Multi-way Analysis
Multi-way Analysis is concerned with data that is measured as a function of three or more factors.
5
Multivariate Images A data array of dimension three (or more) where the first two dimensions are spatial and the last dimension(s) is a function of another variable.
6
Information Hierarchy Data Information Chemistry and Physics
Knowledge Understanding
7
Why Chemometrics? • It’s a multivariate world! •
Need windows into this multivariate world
• There are many things that simply can’t be done if you don’t recognize this, including • • • •
sample classification/pattern recognition calibrations for complex systems (often spectroscopy) transfer of calibrations between instruments fault and upset detection
• Chemometrics focuses on the part of math and statistics applicable to chemical problems • More expensive to do things with hardware if you can do them with math instead 8
Tools of the Trade
9
Principal Components Analysis First PC
8
6
4
Second PC
0 6
Va
riab 4
Variable 3
2
le 2
Mean Vector
2
4
2 0
10
6
0
1 i ble
PCA Math samples
variables
X
=
p1 t1
+
p2 t2
+...+ tk
pk +
E
The pi are the eigenvectors of the covariance matrix T X X cov(X ) = m−1
cov(X )p i = λ i p i and the λi are the eigenvalues. Amount of variance captured by tipi proportional to λi.
11
City Streets Analogy Space Needle
City Streets Analogy Chief Seattle Statue
N Mama’s Mexican Kitchen
Puget Sound
12
© DeLorme 1996 Street Atlas USA ® 3.0 for Macintosh (800)452-5931 © © © © ©DeLorme DeLorme DeLorme DeLorme DeLorme1996 1996 1996 1996 1996Street Street Street Street StreetAtlas Atlas Atlas Atlas AtlasUSA USA USA USA USA® ® ® ® ®3.0 3.0 3.0 3.0 3.0for for for for forMacintosh Macintosh Macintosh Macintosh Macintosh(800)452-5931 (800)452-5931 (800)452-5931 (800)452-5931 (800)452-5931 © DeLorme 1996 Street Atlas USA ® 3.0 for Macintosh (800)452-5931
Properties of PCA • ti,pi pairs ordered by amount of variance captured • variance = information • ti or scores form an orthogonal set Tk which describe relationship between samples • pi or loadings form an orthonormal set Pk which describe relationship between variables
13
Geometry of Q and T2
Variable 3
8
Sample with large Q Unusual variation outside the model
First PC
6
Sample with large T2 Unusual variation inside the model
4
Second PC
2
0 6
Va
14
6
4
riab le 2
4
2
2 0
0
ble 1 a i r a V
PCA Statistics Control limits can be developed for the lack of model fit statistic Q: Q i = e i e Ti = x i (I − P k P Tk )x Ti and Hotelling’s T2 statistic: T 2i = t i λ − 1 t Ti = x i P k λ − 1 P k x Ti Control limits can also be developed for the individual scores (tij) and the residuals (eij)
15
Dirty T-shirt Analogy
PCA attempts to partition data into deterministic and non-deterministic portions PCA
Data
16
Applying a PCA Model to New Data • A PCA model is a description of a data set, including its mean, amount of variance and its direction, dimensionality, and typical residuals • New data can be compared with existing PCA models to see if it is “similar” • Used in Multivariate Statistical Process Control (MSPC)
17
SIMCA x2
Class 1 3 PCs
Class 2 1 PC
Class 3 2 PCs x3
18
x1
Regression • Often want to obtain a relationship between one set of variables, X, and another, y or Y. • Absorbances -> concentrations or other property • Acoustic signature -> particle size distribution
• • • • •
Want y = Xb + e (or Y = XB + E) Relationship may be non-causal May have more variables than samples Highly collinear data Problem if using MLR! 19
Estimation of b: MLR • It is possible to estimate b from b = X +y where X+ is psuedo-inverse of X • There are many ways to obtain a pseudo-inverse, most obvious is Multiple Linear Regression (MLR), a.k.a. Ordinary Least Squares (OLS) • In this case, X+ defined by: X+ = (XTX)-1XT
20
Problem with MLR • Matrix inverse exists only if • • •
Rank(X) = number of variables, but rank(X) ≤ min {mx,nx} X has more samples than variables (problem with spectra) Columns of X are not collinear
• Matrix inverse may exist but be highly unstable if X is nearly rank deficient • Much of multivariate calibration involves tricks for obtaining regression models in spite of problems with matrix inverses!
21
Getting Around the MLR Problem • MLR doesn’t work when mx < nx, or when variables are colinear • Possible solution: eliminate variables, e.g. stepwise regression or other variable selection • •
how to choose which variables to keep? lose multivariate advantage - signal averaging
• Another solution: use PCA to reduce original variables to some smaller number of factors • •
22
retains multivariate advantage noise reduction aspects of PCA
Principal Components Regression • PCR is one way to deal with ill-conditioned regression problems. • Property of interest y is regressed on PCA scores: X+ = Pk(TkTkT)-1TkT • Problem is to determine k, the number of PCs to retain in formation of X+
23
Determining the Number of Factors (PCs or LVs) • A central idea in PCR (and PLS) is that variance is important: use factors that describe lots of variance first • Question: when do you stop? • Answer: use cross-validation • Build model on part of the data and use remaining data to test model as a function of number of factors retained
24
Model Cross-validation and Validation • Cross-validation is a common step in model building • Models should also be validated on totally separate data sets if possible • Why is this important? • It is very easy to fit data, but making predictions is hard!
25
Problem with PCR • Some PCs not relevant for prediction, only relevant for describing X • Result of determining PCs without regard to property to be predicted • Solution: find factors using some information from y (or Y), not just X
26
Solution: Partial Least Squares Regression (PLS) • PLS is related to PCR and MLR • • •
PCR captures maximum variance X MLR achieves maximum correlation with y PLS tries to do both, maximizes covariance
• PLS requires addition of weights W to maintain orthogonal scores • Factors calculated sequentially by projecting y through X • Matrix inverse is: X+ = Wk(PkTWk)-1(TkTTk)-1TkT
27
Cross-validation PRESS Curve 3.9
Note: no irrelevant factors
PRESS
3.8
Best choice for number of LVs
3.7
3.6
3.5
28
0
5
10 Number of Latent Variables
15
20
PLS2 Modelling X-Block Outer Model w1
Y-Block Outer Model 1st PC
1st PC q1
X2
Y2
X1
Inner Model
w1 and q1 are similar to u1 first PCs in X and Y but are rotated so that there is better correlation between the X scores t1 (= Xw1) and Y scores u1 (= Yq1)
29
Y1
Slope = b
t1
Multivariate Curve Resolution • MCR attempts to extract pure component spectra and concentration profiles evolving systems like GC-MS • Given a response matrix Nm that is the product of concentration profiles C and pure component spectra S: Nm = CS + E • Uses alternating and constrained least squares to get C and S
30
The PARAFAC Model c1
D
=
c2
+ b1
a1
31
a2
+ ... + b2
E
Opportunities in Process Analytical Technology (PAT)
32
2
2
1
1 Variable X2
Variable X2
Multivariate Statistical Process Control 0 -1 -2 -4
0
-2
0 Variable X1
2
X1 with 95% Confidence limits
40 60
-2
0 Variable X1
2
0
1 Score on Second PC
Time
-2
4
80
33
0 -1
20
100 -4
X2 with 95% Confidence Limts
4
50 Time
100
Principal Component Scores
0.5 0 -0.5 -1 -1.5 -2 -4
-2 0 2 Score on First PC
4
Example from Distillation • • • • • • • •
41 stage column hexane and isopropanol LV control of top and bottom VT compositions top and bottom controlled to 99% purity full dynamic non-linear simulation noise on temps 0.2 C build PCA model on 100 F, zF normal samples load columMSPCdat
V
PC LC
L
D, xD
LC
B, xB
34
Fault #1: Temperature Sensor Ramped bias (0.2 to 2 C) is added to temperature from tray 35 at sample 31 Scores Plot
8
7
7
6 Q Residual Contribution
Q Residual
6 5 4 3 2
5 4 3 2 1 0 -1
1 0
Sample 43 Q Residual = 64.9764
8
-2 0
10
35
20 30 Sample Number
40
50
-3
0
10
20 Variable Number
30
40
Fault #2: Feed Quality Amount of feed entering as vapor goes from 0% to 50% at time 31 Scores Plot 80
37
2
70
Q Residual Contribution
36
2
60 Hotelling T
Sample 35 Q Residual = 215.1541
3
38
50 35
40 30
39
20 34
10
3 18 19 24 2527 30 17 23 2628 16 20 22 15 21 2931 32 14 33 0 12 45678910 12 13 11 0 10 20 30 Sample Number
36
40 44 45 43 4648 50 41 42 474951 40 50
1 0 -1 -2 -3 -4 -5
60
-6
0
10
20 Variable Number
30
40
TOF-SIMS of Time Release Drug Delivery System • Multilayer drug beads serve as controlled-release delivery system • TOF-SIMS taken of cross section of bead • Evaluate integrity of layers, distribution of ingredients • Thanks again to Physical Electronics and Anna Belu for the data! Reference: A.M. Belu, M.C. Davies, J.M. Newton and N. Patel, “TOF-SIMS Characterization and Imaging of Controlled-Release Drug Delivery Systems, Anal. Chem., 72(22), pps 5625-5638, 2000
37
Total Ion Image of Bead 350
300
250
200
150
100
50
0
38
False Color Image based on Scores of First 3 PCs False Color Image of First 3
39
Inferential Measurements Feedback To Operator for Quality Control and Model Maintenance
Controls
Process Process
Sample Sample
Slow & Infrequent
Lab Lab Analysis Analysis
Auxiliary Variables
Model Model To APC Controller
Inferential Sensor
To Operator for Quality Control (e.g. SPC or MSPC)
40
Inferred Property Feedback Rapid & Continuous
Physical Property
Process Data Characteristics
correlated: variables are not independent
x2 x1
time
autocorrelated: variables correlate with themselves over time
x time crosscorrelated: variables correlate with other variables at different time lags
x2 x1 41
time
Distillation Column • • • • • • •
41 stage column hexane and isopropanol LV control of top and bottom VT compositions top and bottom controlled to 99% purity full dynamic non-linear simulation noise on temps 0.2 C load DIST_EX_2 F, zF
V
PC LC
L
D, xD
LC
B, xB
42
Goal • Develop inferential sensor to predict distillate composition based on tray temperatures • Make model work over a range of operating conditions • Used designed experiment to generate data for identification of model • Can use model for control and/or monitoring purposes
43
Designed Experiment Inputs to Distillation Column
uyDs
99.2 99 98.8 0
200
400
600
800
1000
1200
1400
1600
0
200
400
600
800
1000
1200
1400
1600
0
200
400
600
800 Time
1000
1200
1400
1600
uxBs
99.2 99 98.8
60
uzF
55 50 45 40
44
If Disturbances are Included in Modeling Data, Model Works Model Built on 41 Temperatures
Predicted and Actual Top Composition
0.994
0.992
0.99
0.988
0.986
0.984
45
0
200
400
600
800
Time
1000
1200
1400
1600
Batch MSPC • Multi-way methods can be used to monitor batches • Build PARAFAC or PARAFAC2 model on normal data, apply to new batches • Example from semiconductor etch process • Problem: batches often of unequal length!
46
PARAFAC2 Model The direct fitted PARAFAC2 model is:
Xk = FkDkAT + E X1
subject to constraint that all FkTFk are equal. This is equivalent to the model
Xk = PkFDkAT + E where the Pk are orthonormal
47
20
48 0 2
40
4 6 8 10 12 14 16 18
Vat Valve
TCP Load
TCP Rfl Pwr
TCP Top Pwr
TCP Impedance
TCP Phase Err
TCP Tuner
RF Impedance
RF Pwr
RF Phase Err
RF Load
RF Tuner
Pressure
He Press
Endpt A
RF Btm Rfl Pwr
RF Btm Pwr
Cl2 Flow
BCl3 Flow
PARAFAC2 Contributions
0
-20
-40
-60 20
PARAFAC2 Loadings in Time Mode on New Batches Loadings for Time Dimension Showing Normal Data Range and Fault Wafers
Factor 1 Loadings
60 40 20 0 -20 -40
0
20
40
0
20
40
60
80
100
120
60 80 Sample Number (Time)
100
120
Factor 2 Loadings
40 20 0 -20 -40
49
Summary • Chemometric tools emphasize • Interpretability • Predictive power
• Many places to use these tools in PAT • MSPC, BSPC • Calibrations, inferentials • Analysis of products
50
Contact Information Eigenvector Research, Inc. 830 Wapato Lake Road Manson, WA 98831 Phone: (509)687-2022 Fax: (509)687-7033 Email:
[email protected] Web: eigenvector.com
This document may be downloaded from http://www.eigenvector.com/Docs/Wise_PAT.pdf 52